Research On An Image Distortion Type Based Image Quality Assessment Scheme

Digital Image is commonly a kind of multimedia means for the informationcommunication. During the procedure of collecting, processing, storage andtransmission, images inevitably show the different type and level of distortion becauseof the imperfect devices and technology, so end-users obtain degraded imagery andget incomplete information. Therefore, it has been an important research to build aperfect image quality assessment system that can reasonably and effectively measurethe image quality and further get high quality images. Although subjective assessmentmay accurately reflect the image quality of human vision system, it could hardly beused in real-life applications because of various complicacies and dependencies ofdifferent experimental conditions, such as lighting, color resolution and observers’preferences. In the past decades years, researchers proposed several methods andmodels to evaluate the image quality so that they can be agree with human perceptionas perfectly as possible, and can be applied in the real-time system. In this thesis, anew image quality assessment scheme is proposed based on the distortion typedetection algorithm.Given the fact that image distortion can be introduced by different facts, aconventional IQA algorithm based on a single criterion may not be capable ofperfectly predicting the image quality of different distorted images. Differentperformance of an IQA metric can be obtained from the variation distortion type, anddifferent IQA metric for a distortion type can obtain the different performance. In thisthesis, a frequency domain feature extraction algorithm is proposed firstly to extractthe distortion information; then a classifier of Gaussian Blur, Gaussian White Noise,JPEG Compression and JPEG2000Compression as well as the support vectormachines are employed to classify the distorted images based on the extractedfeatures; By Analyzing the correlation of19IQAs and selecting four IQAs proposedwith different theory with low correlation, we build a heterogeneous image qualityassessment scheme based on multiple linear regressions analysis for each particulardistortion type. After the classification of the image distortion type, correspondingimage quality assessment scheme can be selected automatically and get the objectivescore. Extensive experimentation on the standard image databases (TID2008database,LIVE database and CSIQ database) is carried out, and it is found that the proposedSVM and particle swarm optimization based classifier can relatively perfectly classifythe image distortion type, and the proposed H-IQA scheme provides a superiorperformance of the IQA in terms of mean opinion score (MOS) of subjectiveassessment and can be applied in the real-time application.